SHTA: Semantic Hard Token Correction and Center Alignment for Semi-Supervised Medical Image Segmentation

Xiaofeng Liu; Yiheng Zhong; Zhuoru Zhang; Zimu Zhang

REVIEW 3 major objections 5 minor 22 references

SHTA: Semantic Hard Token Correction and Center Alignment for Semi-Supervised Medical Image Segmentation

T0 review · 3 major / 5 minor · reviewed 2026-07-09 · glm-5.2

Challenge this review Re-run · record.json Download PDF Read on arXiv ↗

Pith's one-line read Fixing organ segmentation at the representation level, not the prediction level

desk verdict Training-only semantic branch for semi-supervised medical image segmentation that shows clear gains on weak baselines but sub-noise improvements on strong ones, with no significance testing. read the letter →

arxiv 2607.07019 v1 pith:OZ33QMJS submitted 2026-07-08 cs.CV

Zhuoru Zhang , Yiheng Zhong , Zimu Zhang , Xiaofeng Liu This is my paper

classification cs.CV

keywords semantichardsegmentationshtasupervisionalignmentcenterconsistency

verification ladder T0 review T1 audit T2 compute T3 formal

The pith

A machine-rendered reading of the paper's core claim, the machinery that carries it, and where it could break.

The reading

Semi-supervised medical image segmentation methods traditionally focus on improving prediction-level supervision: deciding which pseudo-labels, hard regions, or uncertain samples should guide training. This paper argues that this approach leaves a gap. Even after a difficult region is selected for supervision, the internal feature representation of that region may still be ambiguous, sitting between competing organ classes in the model's learned embedding space. The paper proposes SHTA, a training-only auxiliary branch that operates on intermediate feature tokens rather than final predictions. SHTA performs three operations: it organizes tokens into class-aware assignments using learnable proxies, identifies and corrects hard tokens whose assignments conflict with labeled guidance, and aligns the corrected tokens into stable class-level centers. The branch is removed at inference time, so the deployed model is architecturally identical to the baseline. The paper claims that this representation-level correction recovers weak and ambiguous organs that prediction-level supervision alone fails to stabilize, and demonstrates consistent improvements across four semi-supervised frameworks on two abdominal CT datasets.

What carries the argument

SHTA attaches to the decoder feature map of an existing segmentation network. It projects features into a token sequence, computes cosine similarity to learnable class proxies to produce token-to-class assignments, selects hard foreground tokens based on confidence and purity thresholds, corrects those tokens toward their dominant ground-truth class via cross-entropy, and aggregates corrected tokens into class centers that are aligned to ground-truth-derived references via cosine similarity. The entire branch contributes only an auxiliary loss during training and is removed at inference.

What would settle it

A direct falsifier would be a scenario where SHTA's three-stage correction is applied but produces no measurable improvement in semantic consistency metrics or segmentation accuracy over the baseline, particularly on the weak and ambiguous organs it claims to recover. A stronger falsifier would show that the token-level class proportions derived from labeled masks are themselves unreliable as semantic references, such that the correction targets are no better than random assignment.

Watch

Extended reading notes

Core claim

The central claim is that post-selection semantic ambiguity, not just prediction noise, is a limiting factor in semi-supervised medical image segmentation, and that this ambiguity can be reduced by converting labeled masks into token-level semantic guidance that directly corrects and aligns the intermediate representations of hard regions. The mechanism is a three-stage pipeline: proxy-based semantic assignment, hard-token refinement using dominant-class targets from labeled tokens, and class-center alignment against ground-truth-derived references. The paper shows that under matched hard-token supply, adding the semantic scaffold and center alignment reduces the correction loss, supporting.

Load-bearing premise

The method assumes that converting labeled masks into token-level class proportions provides a valid semantic reference for correcting hard-token assignments. If the spatial resolution of tokenization is too coarse relative to the organ boundaries, the dominant foreground class assigned to a token may be noisy or incorrect, which would undermine the correction signal.

Editorial extensions

If this is right

If representation-level correction is the bottleneck after region selection, then existing semi-supervised methods that invest heavily in pseudo-label refinement may be hitting diminishing returns because they are not addressing the embedding-level ambiguity that SHTA targets.
The training-only, inference-free design means that deployed clinical segmentation models could be improved without any change to their runtime architecture, which matters for deployment in memory-constrained medical imaging pipelines.
The token-level class-proportion guidance derived from labeled masks could be applied to other dense prediction tasks where class imbalance and boundary ambiguity persist, such as histology or satellite imagery.
The finding that weak-organ recovery is driven by semantic correction rather than token quantity suggests that future work should focus on the quality of representation-level supervision rather than the volume of selected hard evidence.

Share X Bluesky LinkedIn Reddit HN

Editorial analysis

A structured set of objections, weighed in public.

Desk editor's note, referee report, and a circularity audit.

Desk Editor's Note

read the letter

The paper to know about: SHTA adds a training-only semantic representation branch to existing semi-supervised segmentation frameworks. The core idea is that after you select hard regions for supervision, the token-level representations in those regions are still semantically ambiguous—so you should correct them at the representation level, not just at the prediction level. The branch does three things: proxy-based semantic assignment, hard-token refinement using labeled masks converted to token-level class proportions, and class-center alignment. It's removed at inference, so zero deployment cost. Code is reportedly included (anonymous repo link in the abstract, which is a good sign for reproducibility intent, though I haven't verified it runs). The framing of 'post-selection semantic ambiguity' is genuinely useful—it names a real gap in the SSL medical segmentation literature, where most work stops at selecting which regions or pseudo-labels to supervise and doesn't touch the intermediate representation geometry. The qualitative analysis in Figures 5-7, showing token reassignment flows and entropy reduction, is the kind of mechanistic evidence I like seeing. The ablation in Table III is clean: each component contributes something different, and the matched-hard-token control in Table IV is a nice touch showing the gains come from better correction, not more selected tokens. The weak-organ recovery on URPC is real and large (RAG 0→38.47 on Synapse). That's the strongest result in the paper. Here's the problem: the claim is 'consistent paired improvements across frameworks,' but for the strong baselines the Dice gains are tiny. MagicNet: +0.12 on Synapse, +0.11 on AMOS. CPS on AMOS: +0.27. In 3D medical segmentation, run-to-run variance routinely hits ±0.3-0.5 Dice. No standard deviations, no multi-seed runs, no significance tests. If those sub-0.3 gains are noise, the 'consistent' qualifier fails for exactly the setting where they claim SHTA refines strong baselines. The ASD regressions (MagicNet +0.60 on Synapse, CPS +0.74 on AMOS) are acknowledged but undercut the consistency narrative. The reader's concern about Y_tok validity is real but secondary—the purity and foreground thresholds partially mitigate coarse-token mixing, and the correction targets come from labeled GT. The bigger issue is statistical. The hyperparameter count is high (lambda_shta, tau_sa, four alpha weights, three thresholds, selection ratio r), and none are stated in the paper body. That's a reproducibility concern. This is a solid, well-motivated method paper with honest reporting (they didn't hide the ASD regressions) but insufficient statistical rigor for the central claim. It deserves a serious referee who will push for multi-seed validation and hyperparameter disclosure. The weak-organ recovery results alone justify the review.

Referee Report

3 major / 5 minor

Summary. This manuscript proposes SHTA, a training-only auxiliary semantic representation branch for semi-supervised medical image segmentation. The method addresses 'post-selection semantic ambiguity' by operating at the representation level rather than the prediction level. SHTA consists of three components: Semantic Assignment (organizing token embeddings via learnable class proxies), Hard Token Refinement (correcting ambiguous hard-token assignments using labeled guidance), and Semantic Center Alignment (stabilizing class-level representation geometry). The branch is attached during training and removed during inference, incurring no additional deployment cost. The authors integrate SHTA into four representative SSL frameworks (GA-CPS, CPS, URPC, MagicNet) and evaluate on the Synapse (20% labeled) and AMOS (5% labeled) datasets, reporting paired improvements in mean Dice and weak-organ recovery.

Significance. The paper tackles a well-motivated problem: the gap between prediction-level supervision and representation-level semantic consistency in semi-supervised segmentation. The design is plug-and-play, training-only, and preserves the base framework's inference pathway, making it practically appealing for clinical deployment. The provision of anonymized code and the matched-hard-token ablation (Table IV) are commendable, as they allow scrutiny of whether gains stem from mechanism improvement rather than mere token selection inflation. The qualitative semantic consistency analyses (Figs. 5-7) provide useful interpretability of the representation-level corrections.

major comments (3)

Tables I and II report single-run results with no standard deviations, multi-seed runs, or significance tests. For the stronger baselines, the Dice gains are very small: MagicNet improves +0.12 on Synapse (65.86 to 65.98) and +0.11 on AMOS (63.47 to 63.58), while CPS on AMOS gains +0.27 (61.93 to 62.20). In 3D medical image segmentation, run-to-run variance from random initialization, data shuffling, and pseudo-label stochasticity routinely produces fluctuations of similar or larger magnitude. If these sub-0.3 gains fall within run-to-run variance, the claim of 'consistent paired improvements across frameworks' is unverified for 3 of 8 paired comparisons. The authors should provide multi-seed results (at least 3 runs) with standard deviations or paired significance tests to substantiate the central claim.
The hard-token selection in Eq. (7) depends on thresholds T_conf, T_fg, T_pur, and the selection ratio r, and the proxy assignment q_i is itself trained by the SHTA loss (Eq. 5). This introduces a degree of circularity: the module improves the representations it uses to select hard tokens. While Table IV attempts to control for this by matching the hard-token supply (3.76%), it does not address whether the selection quality or proxy drift during training biases the correction targets. The authors should discuss this dependency and ideally report sensitivity analyses for the key thresholds (T_conf, T_fg, T_pur, r) to show robustness.
The claim of 'consistent' improvements is contradicted by ASD regressions in two paired comparisons: MagicNet on Synapse (ASD worsens from 3.08 to 3.68, +0.60) and CPS on AMOS (ASD worsens from 10.83 to 11.57, +0.74). The paper acknowledges this metric dependency in the conclusion, but the abstract and main results framing still use 'consistent paired improvements.' The framing should be tempered to reflect that gains are concentrated in Dice and weak-organ recovery, while boundary-related metrics (ASD) vary by framework and dataset.

minor comments (5)

Section III-A, Eq. (1): The tokenization process P(.) and the conversion of labeled masks to token-level class proportions Y_tok = TokDist(Y_L) are underspecified. It is unclear how spatial resolution of tokens relates to the mask resolution, and whether coarse tokenization could mix classes within a token. A brief clarification of the tokenization scheme and its resolution would help readers assess the validity of the dominant-class target in Eq. (8).
Figure 2: The labels and arrows in the diagram are small and somewhat difficult to parse. In particular, the distinction between the semantic-alignment paths (yellow arrows) and the training flow (solid arrows) could be made clearer, perhaps with a higher-resolution figure or additional subfigure captions.
Table I: The 'Reported SSL' rows include methods (e.g., DCA, SKCDF) that appear to outperform the SHTA-enhanced baselines (e.g., DCA 69.90 vs. GA-CPS+SHTA 68.56). While these may use different architectures or settings, a brief note clarifying why these external references are not directly comparable would contextualize the results.
Section IV-B: The statement 'SHTA can refine a strong prediction-consistency baseline rather than only rescuing weak models' is made for CPS, but the CPS gain on AMOS is only +0.27 Dice with an ASD regression. This characterization may overstate the evidence for strong baselines.
The loss weights (lambda_shta, alpha_as, alpha_a, alpha_h, alpha_c) and temperature tau_sa are mentioned but their values are not specified in the main text. These should be reported, at least in a supplementary table or footnote, for reproducibility.

Circularity Check

0 steps flagged · score 1.0 of 10

No significant circularity found; the derivation chain is self-contained against external ground truth.

full rationale

The paper's derivation chain proceeds as follows: (1) GT labeled masks are converted into token-level class proportions Y_tok via TokDist (Eq. 1), providing an externally-derived semantic reference. (2) Semantic Assignment (Eq. 3-5) introduces learnable proxies and supervises token-to-class assignments against Y_tok, with proxy calibration against GT-derived class references c_k^gt (Eq. 4). (3) Hard Token Refinement (Eq. 6-8) selects hard tokens using a combination of model confidence q_i and GT-derived quantities (foreground ratio, purity), then corrects them toward the dominant GT foreground class y_i^hard = argmax_{k>0} Y_tok_{i,k}. (4) Semantic Center Alignment (Eq. 9-10) aligns corrected token centers with the same GT-derived references. (5) Final evaluation is against held-out test set Dice/ASD scores on Synapse and AMOS benchmarks. The correction targets throughout come from labeled GT, not from the model's own predictions. The hard-token selection does depend on q_i (the model's current assignment), which is trained by the SHTA loss — this creates a minor self-referential loop in training dynamics (the module refines representations it uses to identify what needs refining). However, this is standard in hard-example mining and does not make the final evaluation circular, since: (a) the selection also incorporates GT-derived thresholds, (b) the correction targets are externally defined, and (c) the reported improvements are measured on a held-out test set against external ground truth. Table IV explicitly controls for this by matching hard-token supply (3.76% in both settings) and showing that the correction loss drops from 0.129 to 0.054 when semantic assignment and center alignment are added, ruling out the alternative that SHTA merely selects more tokens. No self-citation chain is load-bearing for the central claim. The skeptic's concern about small effect sizes and missing multi-seed runs is a statistical validity issue, not a circularity issue. Score 1 reflects the minor training-dynamics self-reference in hard-token selection, which is not load-bearing for the central empirical claim.

Assumptions & free parameters 5 free parameters · 2 assumptions · 1 invented entities

The method introduces several hyperparameters and thresholds that are fitted or selected via validation. The core axioms are standard domain assumptions for representation learning but depend on the tokenization resolution being appropriate.

free parameters (5)

lambda_shta
Overall loss weight for SHTA branch, selected by validation.
tau_sa
Semantic-assignment temperature.
alpha_as, alpha_a, alpha_h, alpha_c
Loss weights for the four SHTA sub-losses.
T_conf, T_fg, T_pur
Thresholds for confidence, foreground ratio, and purity used in hard-token selection (Eq. 7).
r
Hard-token selection ratio (Top-r).

assumptions (2)

domain assumption Token-level class proportions derived from labeled masks provide a valid semantic reference for representation correction.
The method relies on TokDist(Y_L) converting masks into token proportions (Eq. 1) and using the dominant class as a correction target (Eq. 8). This assumes the tokenization resolution is meaningful.
domain assumption Learnable class proxies can capture stable semantic centers without collapsing.
The Semantic Assignment module (III-B) introduces proxies mu_k and assumes they can be calibrated by labeled references to provide stable assignment targets.

invented entities (1)

Learnable class proxies (mu_k) independent evidence
purpose: To organize token embeddings into class-aware structures in Semantic Assignment.
Standard technique in representation learning; calibrated by labeled data in the paper.

how reviews work

0 comments

Cite this review

Pith. "Pith review of SHTA: Semantic Hard Token Correction and Center Alignment for Semi-Supervised Medical Image Segmentation." pith.science (2026). https://pith.science/paper/OZ33QMJS

@misc{pith2026260707019,
  author       = {Pith},
  title        = {Pith review of: SHTA: Semantic Hard Token Correction and Center Alignment for Semi-Supervised Medical Image Segmentation},
  year         = {2026},
  howpublished = {\url{https://pith.science/paper/OZ33QMJS}},
  note         = {Machine review of arXiv:2607.07019}
}

read the original abstract

Recent advances in semi-supervised medical image segmentation have achieved remarkable performance through prediction consistency, pseudo-label supervision, and hard-region supervision. However, these methods primarily improve supervision quality rather than explicitly enforcing semantic consistency in the learned representations of hard regions. Consequently, even under increasingly stronger prediction-level supervision, difficult regions exhibiting unstable semantic assignment often fail to establish semantically consistent representations during training, thereby limiting further segmentation improvement. To address this issue, we propose SHTA (Semantic Hard Token Correction and Center Alignment), a lightweight training-time semantic representation branch. Instead of introducing additional prediction supervision, SHTA refines intermediate semantic representations through Semantic Assignment, Hard Token Refinement, and Semantic Center Alignment, thereby improving semantic consistency in hard regions while preserving the original prediction pathway and introducing no additional inference cost. We integrate SHTA into representative semi-supervised segmentation frameworks, including GA-CPS, CPS, URPC, and MagicNet, and conduct evaluations on the Synapse and AMOS datasets. Experimental results demonstrate that SHTA delivers consistent paired improvements across frameworks, with especially clear gains in segmentation accuracy, weak-organ recovery, and semantic ambiguity reduction, while incurring only training-time overhead. The code is available at https://anonymous.4open.science/r/release_SHTA-42D5/.

Figures

Figures reproduced from arXiv: 2607.07019 by the authors.

**Figure 2.** Overview of the SHTA semantic branch. SHTA contains (b) Semantic Assignment, (c) Hard Token Refinement, and (d) Semantic Center Alignment. [PITH_FULL_IMAGE:figures/full_fig_p003_2.png] view at source ↗

**Figure 3.** Training and inference overhead ratios across different frameworks, [PITH_FULL_IMAGE:figures/full_fig_p005_3.png] view at source ↗

Figures from the paper (3 more)

**Figure 4.** Figure 4: Qualitative comparison across SSL frameworks on Synapse. B and F denote baseline and full SHTA variants; red contours denote GT boundaries and [PITH_FULL_IMAGE:figures/full_fig_p007_4.png]

**Figure 5.** Figure 5: Region-level ambiguity reduction on Synapse. Zoomed regions compare GT, baseline predictions, SHTA predictions, and entropy maps; arrows mark [PITH_FULL_IMAGE:figures/full_fig_p007_5.png]

**Figure 6.** Figure 6: Representation-level semantic consistency on Synapse. Decoder [PITH_FULL_IMAGE:figures/full_fig_p007_6.png]

Discussion (0). Sign in to comment.

Reference graph

Works this paper leans on

22 extracted references · 22 canonical work pages

[1]

3D U-Net: Learning dense volumetric segmentation from sparse annotation,

O. Cicek, A. Abdulkadir, S. S. Lienkamp, T. Brox, and O. Ron- neberger, “3D U-Net: Learning dense volumetric segmentation from sparse annotation,” inMedical Image Computing and Computer-Assisted Intervention – MICCAI 2016, LNCS, vol. 9901, 2016, pp. 424–432

work page 2016
[2]

nnU-Net: a self-configuring method for deep learning-based biomedical image segmentation,

F. Isensee, P. F. Jaeger, S. A. A. Kohl, J. Petersen, and K. H. Maier-Hein, “nnU-Net: a self-configuring method for deep learning-based biomedical image segmentation,”Nature Methods, vol. 18, no. 2, pp. 203–211, 2021

work page 2021
[3]

PG-SAM: A fine-grained prior-guided SAM framework for prompt-free medical image segmentation,

Y . Zhong, Z. Luo, C. Liu, F. Tang, Y . Hu, Z. Peng, M. Hu, J. Su, Z. Ge, and I. Razzak, “PG-SAM: A fine-grained prior-guided SAM framework for prompt-free medical image segmentation,” inProceedings of the 2025 IEEE International Conference on Bioinformatics and Biomedicine, 2025, pp. 3369–3376

work page 2025
[4]

TransUNet: Rethinking the U-Net architecture design for medical image segmentation through the lens of transformers,

J. Chen, J. Mei, X. Li, Y . Lu, Q. Yu, Q. Wei, X. Luo, Y . Xie, E. Adeli, Y . Wang, M. P. Lungren, S. Zhang, L. Xing, L. Lu, A. Yuille, and Y . Zhou, “TransUNet: Rethinking the U-Net architecture design for medical image segmentation through the lens of transformers,”Medical Image Analysis, vol. 97, p. 103280, 2024

work page 2024
[5]

Uncertainty-aware self-ensembling model for semi-supervised 3D left atrium segmenta- tion,

L. Yu, S. Wang, X. Li, C.-W. Fu, and P.-A. Heng, “Uncertainty-aware self-ensembling model for semi-supervised 3D left atrium segmenta- tion,” inMedical Image Computing and Computer Assisted Intervention, 2019, pp. 605–613

work page 2019
[6]

Semi-supervised medical image segmentation via uncertainty rectified pyramid consistency,

X. Luo, G. Wang, W. Liao, J. Chen, T. Song, Y . Chen, S. Zhang, D. N. Metaxas, and S. Zhang, “Semi-supervised medical image segmentation via uncertainty rectified pyramid consistency,”Medical Image Analysis, vol. 80, p. 102517, 2022

work page 2022
[7]

Semi-supervised semantic segmentation with cross pseudo supervision,

X. Chen, Y . Yuan, G. Zeng, and J. Wang, “Semi-supervised semantic segmentation with cross pseudo supervision,” inProceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, 2021, pp. 2613–2622

work page 2021
[8]

Exploring smoothness and class-separation for semi-supervised medical image segmentation,

Y . Wu, Z. Wu, Q. Wu, Z. Ge, and J. Cai, “Exploring smoothness and class-separation for semi-supervised medical image segmentation,” Medical Image Analysis, vol. 90, p. 102976, 2023

work page 2023

Show all 22 references

[9]

Bidirectional copy-paste for semi-supervised medical image segmentation,

Y . Bai, D. Chen, Q. Li, W. Shen, and Y . Wang, “Bidirectional copy-paste for semi-supervised medical image segmentation,” inProceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, 2023, pp. 11514–11524

2023
[10]

MagicNet: Semi-supervised multi-organ segmentation via magic-cube partition and recovery,

D. Chen, Y . Bai, W. Shen, Q. Li, L. Yu, and Y . Wang, “MagicNet: Semi-supervised multi-organ segmentation via magic-cube partition and recovery,” inProceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, 2023, pp. 23869–23878

2023
[11]

Adaptive bidirectional displace- ment for semi-supervised medical image segmentation,

H. Chi, J. Pang, B. Zhang, and W. Liu, “Adaptive bidirectional displace- ment for semi-supervised medical image segmentation,” inProceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recogni- tion, 2024, pp. 4070–4080

2024
[12]

Dual-debiased heterogeneous co-training frame- work for class-imbalanced semi-supervised medical image segmenta- tion,

H. Wang and X. Li, “Dual-debiased heterogeneous co-training frame- work for class-imbalanced semi-supervised medical image segmenta- tion,” inMedical Image Computing and Computer Assisted Intervention, 2023, pp. 582–591

2023
[13]

A semantic knowledge complementarity based decoupling framework for semi-supervised class-imbalanced medical image segmentation,

Z. Zhang, G. Yin, B. Zhang, W. Liu, X. Zhou, and W. Wang, “A semantic knowledge complementarity based decoupling framework for semi-supervised class-imbalanced medical image segmentation,” inPro- ceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, 2...

2025
[14]

Gradient-aware for class-imbalanced semi-supervised medical image segmentation,

W. Qi, J. Wu, and S. C. Chan, “Gradient-aware for class-imbalanced semi-supervised medical image segmentation,” inEuropean Conference on Computer Vision, 2024, pp. 473–490

2024
[15]

Divide, conquer, and aggregate: Asymmetric experts for class- imbalanced semi-supervised medical image segmentation,

Y . Liu, “Divide, conquer, and aggregate: Asymmetric experts for class- imbalanced semi-supervised medical image segmentation,” inProceed- ings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, 2025, pp. 8503–8513

2025
[16]

All-around real label supervision: Cyclic prototype consistency learning for semi-supervised medical image segmentation,

Z. Xu, Y . Wang, D. Lu, L. Yu, J. Yan, J. Luo, K. Ma, Y . Zheng, and R. K.-y. Tong, “All-around real label supervision: Cyclic prototype consistency learning for semi-supervised medical image segmentation,” arXiv preprint arXiv:2109.13930, 2021

2021 arXiv
[17]

Self-aware and cross-sample prototypical learning for semi-supervised medical image segmentation,

Z. Zhang, R. Ran, C. Tian, H. Zhou, X. Li, F. Yang, and Z. Jiao, “Self-aware and cross-sample prototypical learning for semi-supervised medical image segmentation,”arXiv preprint arXiv:2305.16214, 2023

2023 arXiv
[18]

Mixed prototype consistency learning for semi-supervised med- ical image segmentation,

L. Li, “Mixed prototype consistency learning for semi-supervised med- ical image segmentation,”arXiv preprint arXiv:2404.10717, 2024

2024 arXiv
[19]

Prototype contrastive consis- tency learning for semi-supervised medical image segmentation,

S. He, Z. Lai, R. Wang, and H. Kong, “Prototype contrastive consis- tency learning for semi-supervised medical image segmentation,”arXiv preprint arXiv:2502.06650, 2025

2025 arXiv
[20]

Inherent consistent learning for accurate semi-supervised medical image segmentation,

Y . Zhu, J. Yang, S.-Q. Liu, and R. Zhang, “Inherent consistent learning for accurate semi-supervised medical image segmentation,” inProceed- ings of Machine Learning Research, vol. 227, 2023, pp. 1581–1601

2023
[21]

Multi-atlas labeling beyond the cranial vault,

B. Landman, Z. Xu, J. E. Iglesias, M. Styner, T. R. Langerak, and A. Klein, “Multi-atlas labeling beyond the cranial vault,” inMICCAI Workshop on Multi-Atlas Labeling, 2015

2015
[22]

AMOS: A large-scale abdominal multi- organ benchmark for versatile medical image segmentation,

Y . Ji, H. Bai, J. Yang, C. Ge, Y . Zhu, R. Zhang, Z. Li, L. Zhang, W. Ma, X. Wan, and P. Luo, “AMOS: A large-scale abdominal multi- organ benchmark for versatile medical image segmentation,”Advances in Neural Information Processing Systems, vol. 35, pp. 36722–36732, 2022

2022

Pith tools

Reviewed July 9, 2026 · model on record in the stance chip above.

[1] [1]

3D U-Net: Learning dense volumetric segmentation from sparse annotation,

O. Cicek, A. Abdulkadir, S. S. Lienkamp, T. Brox, and O. Ron- neberger, “3D U-Net: Learning dense volumetric segmentation from sparse annotation,” inMedical Image Computing and Computer-Assisted Intervention – MICCAI 2016, LNCS, vol. 9901, 2016, pp. 424–432

work page 2016

[2] [2]

nnU-Net: a self-configuring method for deep learning-based biomedical image segmentation,

F. Isensee, P. F. Jaeger, S. A. A. Kohl, J. Petersen, and K. H. Maier-Hein, “nnU-Net: a self-configuring method for deep learning-based biomedical image segmentation,”Nature Methods, vol. 18, no. 2, pp. 203–211, 2021

work page 2021

[3] [3]

PG-SAM: A fine-grained prior-guided SAM framework for prompt-free medical image segmentation,

Y . Zhong, Z. Luo, C. Liu, F. Tang, Y . Hu, Z. Peng, M. Hu, J. Su, Z. Ge, and I. Razzak, “PG-SAM: A fine-grained prior-guided SAM framework for prompt-free medical image segmentation,” inProceedings of the 2025 IEEE International Conference on Bioinformatics and Biomedicine, 2025, pp. 3369–3376

work page 2025

[4] [4]

TransUNet: Rethinking the U-Net architecture design for medical image segmentation through the lens of transformers,

J. Chen, J. Mei, X. Li, Y . Lu, Q. Yu, Q. Wei, X. Luo, Y . Xie, E. Adeli, Y . Wang, M. P. Lungren, S. Zhang, L. Xing, L. Lu, A. Yuille, and Y . Zhou, “TransUNet: Rethinking the U-Net architecture design for medical image segmentation through the lens of transformers,”Medical Image Analysis, vol. 97, p. 103280, 2024

work page 2024

[5] [5]

Uncertainty-aware self-ensembling model for semi-supervised 3D left atrium segmenta- tion,

L. Yu, S. Wang, X. Li, C.-W. Fu, and P.-A. Heng, “Uncertainty-aware self-ensembling model for semi-supervised 3D left atrium segmenta- tion,” inMedical Image Computing and Computer Assisted Intervention, 2019, pp. 605–613

work page 2019

[6] [6]

Semi-supervised medical image segmentation via uncertainty rectified pyramid consistency,

X. Luo, G. Wang, W. Liao, J. Chen, T. Song, Y . Chen, S. Zhang, D. N. Metaxas, and S. Zhang, “Semi-supervised medical image segmentation via uncertainty rectified pyramid consistency,”Medical Image Analysis, vol. 80, p. 102517, 2022

work page 2022

[7] [7]

Semi-supervised semantic segmentation with cross pseudo supervision,

X. Chen, Y . Yuan, G. Zeng, and J. Wang, “Semi-supervised semantic segmentation with cross pseudo supervision,” inProceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, 2021, pp. 2613–2622

work page 2021

[8] [8]

Exploring smoothness and class-separation for semi-supervised medical image segmentation,

Y . Wu, Z. Wu, Q. Wu, Z. Ge, and J. Cai, “Exploring smoothness and class-separation for semi-supervised medical image segmentation,” Medical Image Analysis, vol. 90, p. 102976, 2023

work page 2023